webAI co-founder and CEO David Stout (@Davidstout) released a pair of small formal-logic models on August 10th, extending his six-year bet that useful enterprise AI will run near private data instead of inside a distant cloud.
The TwiL-LM release includes 1.7 billion and 3 billion parameter variants designed to translate plain English into formal logic, determine whether conclusions follow from premises and handle multi-step deduction. webAI says both are available through Hugging Face for Transformers and llama.cpp under webAI's non-commercial model license.
Stout grew up on a Michigan ranch before moving to Austin, according to an Austin Business Journal profile. He founded webAI in 2019 with Ethan Baird, now chief research officer, and Tyler Mauer (@TylerMauer4), webAI's chief product officer. webAI describes the original team as Michigan engineers who concluded that sensitive AI workloads needed to run locally.
TwiL turns that founding view into a narrowly defined product. Rather than trying to answer every kind of prompt, the models concentrate on structured tasks behind compliance policies, contract conditions, theorem statements and rule-based decisions. In the release, Stout describes the model as "almost like an auto-correct for AI," a reasoning layer that can inspect and restructure output from other specialized models.
The benchmark claim
webAI's headline comparison pits TwiL-LM3, the 3B variant, against OpenAI's gpt-oss-120b. In webAI's formal-reasoning suite, TwiL-LM3 scored higher on four of five reported tests: rule induction, semantic parsing, Lean formalization and exact-format answering. gpt-oss-120b retained the lead on entailment labeling, scoring 77.5 against TwiL-LM3's 68.7.
The largest reported gaps came in semantic parsing, where webAI measured 87.6 for TwiL-LM3 against 43.3 for gpt-oss-120b, and exact-format answering, where the scores were 52.0 and 7.0. webAI also says TwiL-LM3 produced 32.9 answers per second in its throughput test, compared with 12.6 for the 120B model.
Those numbers measure performance on webAI's own suite. They have not been independently reproduced, and the model card says the tests should not be read as evidence of general intelligence. The comparison also covers models with different architectures, training methods, output limits and parameter counts, which makes it useful as a product demonstration rather than a controlled scaling study.
The detailed model card for the 1.7B version is unusually direct about those limits. webAI reports that the smaller model improved its SmolLM2-1.7B base from 0.185 to 0.361 on webAI's aggregate formal-logic score. Outside that specialty, the gains were mixed: the adapter improved LogicBench performance while scoring below the base model on GSM8K and two ARC-Challenge evaluations.
webAI also warns that TwiL is not a verifier. Formal expressions can still be syntactically valid and logically wrong, so deployments involving legal, medical, financial or safety-critical decisions require a symbolic solver, theorem prover or human review.
A 289 MB fine-tune
The smaller TwiL model was built as a LoRA adapter for SmolLM2-1.7B-Instruct rather than trained from scratch. The adapter adds about 72.35 million parameters and occupies roughly 289 MB. webAI says the adapter was trained on a purpose-engineered, proprietary formal-logic data engine built from open sources.
That approach kept the deployable package small. The recommended Q4 quantized build is 1.06 GB, with larger Q5, Q8 and FP16 versions also listed. The model supports an 8,192-token context window and local llama.cpp deployment. webAI measured aggregate generation throughput of 366.8 tokens per second in its evaluation, though the card cautions that the figure depends on hardware, batching, prompt length, precision and backend configuration.
The training corpus itself is not public, limiting independent reproduction and contamination analysis. The release artifacts also lack optimizer settings, learning-rate schedules, seed data, training hardware and confidence intervals. Those omissions matter because TwiL's strongest sales argument rests on benchmark efficiency: webAI is asking developers to accept that a targeted data engine can substitute for tens of billions of general-purpose parameters on a defined class of work.
Stout's network of experts
TwiL fits the architecture Stout and his co-founders have been describing throughout 2026. In a March webAI presentation, webAI argued that organizations should own networks of specialized models trained around local knowledge. Each model handles a defined task, and the network combines those outputs without sending private data to a metered API.
TwiL gives that network a potential referee. A contract model could extract clauses, a retrieval model could find supporting documents and TwiL could convert the result into formal rules or test whether a conclusion follows. The small footprint allows that checking step to run continuously on local hardware, which is the practical value behind Stout's auto-correct analogy.
Commercial use will require a separate arrangement because webAI released the weights under its Non-Commercial License v1.0. The license lets developers inspect and test the model on Hugging Face, while commercial deployment requires permission from webAI. That structure turns the release into both a research artifact and a potential entry point to webAI's private AI platform.
Stout has substantial capital behind the strategy. webAI said it had raised $60 million by September 2024, when its Series A valued webAI at $700 million. In January 2026, webAI announced another round at a $2.5 billion pre-money valuation, with TIME Ventures, Atreides Management, Forerunner Ventures and OXCART Ventures participating. Axios reported that financing totaled high double-digit millions.
That funding also established the Intelligence Lab led by Paul J. Maykish, the group credited with TwiL. The release is the clearest technical output yet from that investment: a downloadable model that developers can benchmark against their own rules and documents. Its value will depend on performance outside webAI's test harness, especially when formal logic meets long contracts, incomplete policies and inconsistent real-world data.